What are Deep Fakes?: Complete Explanation
People who have watched Mark Zuckerberg talk about how he has total control of data for billions of individuals have watched a deep fake. Jon Snow apologizing for the ending to the Game of Thrones show was also a deep fake. These are just a few examples of deep fakes on the internet in recent years. Deep fakes are usually in video form, but they can be made in audio form as well.
Deep fakes implement a form of artificial intelligence known as deep learning. Deep learning can make computer-generated fake images of people, objects, or events and turn them into what appears to be a real picture or video. If a person wants to make a famous person say specific things or people want to make it look like they are starring in a well-known movie, the technology is now available to accomplish this.
What are Deep Fakes?: An Exact Definition
Technology that includes software, algorithms, and even camera apps has become so advanced that users can now easily alter online images and even create fake videos that look extraordinarily real. What started as the ability to remove facial imperfections, elongate legs, or ad fun animal ears has evolved into what are known as deep fakes.
The software, algorithms, and apps used are produced and run by artificial intelligence. This particular type of artificial intelligence includes a subcategory of AI called deep learning. Deep learning can learn and even make intelligent decisions through an arrangement of different kinds of algorithms. This deep learning “learns” how to copy and reproduce images that are very similar to the original image.
Deep Fakes, just like different types of malware or computer viruses, are often very difficult to detect. They can also, in different ways, be just as destructive and even become security threats. While deep fakes are sometimes used for harmless humor, entertainment purposes, or even videos for educational instruction, there are several reasons why they can also be extremely dangerous. The following are a few examples of the risks of deep fake technology.
- Create Revenge Porn: Placing one person’s head on the body of another individual in a pornographic film or the face of one person over another on a nude picture is sometimes used in what is called revenge porn. It’s estimated by Google that revenge porn gets approximately 95,000 searches each month.
- Create False Narratives: Using deep fakes, cybercriminals could create a variety of videos or images on the internet to make the public believe that another person or group is saying or doing something they aren’t.
- Spread Fear: Having a CEO of a large company or a world leader appears to say something that he or she is not, could potentially have disastrous results.
- Cause Distrust: Cybercriminals are not only involved in extortion, fraud, and identity theft but in causing discord and distrust among the public. Since it’s sometimes difficult to know if what we’re seeing on the internet is true or not, many people may begin to distrust or doubt even things that are true.
How do Deep Fakes Work?
With the right kind of technology, deep fakes are now relatively easy to make. The deep learning process is able to create a believable counterfeit through studying a variety of photographs or videos. AI studies people and events to imitate them from different angles. Eventually, the algorithms learn to mimic the person or event so closely that it’s hard to tell the fake from the real thing.
The Studying Process
The first step involves running hundreds, perhaps thousands, of face shots through an artificial intelligence algorithm that is called an encoder. The encoder will find and learn the details of each of these faces and the similarities between the two people. AI will then reduce all the details in these thousands of shots into shared features that are common or similar. These images are compressed in the process.
The Generator Process
A generative adversarial network, or what is called GAN, is another step in the process to create deep fakes. This method actually pits two AI algorithms against one another. The first algorithm is called the generator. The generator is given random noise that it can turn into an actual image. This is basically a new synthetic image that will be fed into a stream of real images.
The Discriminator Process
The real images are put into a second algorithm that is called a discriminator. When this process first begins, the synthetic images don’t look at all like the real images. When repeating the process over and over, however, the generator and discriminator receive feedback. After enough feedback and cycles, the generator will begin to produce realistic faces of a nonexistent person.
How do You Create Deep Fakes?
Image and video manipulations have been occurring for many years. Swapping faces and putting them on the bodies of celebrities or porn performers is now relatively easy to do. Unfortunately, this is now used in cyberattacks. For example, cybercriminals can convincingly imitate a bank president’s voice through audio or recreate a video featuring the person. A crime is then committed through extortion or what seems to be regular business practices. Whether for malicious attacks or benign purposes, how are deep fakes created?
Deep Fakes previously required complex algorithms such as facial recognition algorithms. Variational auto-encoders were also used to make deep fakes. Variational auto-encoders (VAE) can encode an image that the individual puts into the artificial intelligence algorithm. The encoder then transforms this particular image into a new image.
There are now apps that enable people to create deep fakes more quickly and easily. These apps and encoders are specifically trained to recognize a person’s face and features. A separate app or encoder is also required that has training in diversity of different faces to create a contrast.
After training an encoder to recognize a face, such as Oprah Winfrey as well as other faces, the deep fake will combine Winfrey’s face with information on other faces and create what appears to be her face. This is usually done on another person’s body. When creating an audio deep fake, it’s necessary to add enough background noise to help disguise that the voice is fake.
It’s difficult to make a convincing deep fake on a basic computer. A high-end desktop with powerful graphics or computing technology from the cloud is normally needed. Even with the best technology the processing time still may take hours to complete.
Even advanced technology isn’t enough to make convincing deep fakes. It takes expertise and experience to touch up videos and reduce various visual defects. There are, however, tools and even companies that will help people accomplish these types of details.
The mobile app Zao is an example of one such tool that people can use on their phones. This app allows users to put their faces on several movie and TV characters that the system has already trained.
Deep learning produces a convincing counterfeit in two distinct stages:
- The system studies videos or photos of the targeted image or person from a variety of angles. The algorithms begin to mimic behaviors, gestures, and speech patterns.
- After an initial fake is produced, the second step begins. Generative adversarial networks, or what are called GANS, take over to detect minor flaws. These flaws are then addressed and improved.
Where did Deep Fakes Originate From?
The name deep fake is related to the technology term, “deep learning.” Deep learning is when algorithms have the ability to teach themselves how to solve problems with data. This type of deep learning has the ability to swap faces in digital content to make deep fakes that are realistic and can fool the average person.
Deep fake technology is something that appears relatively new in the 21st century. The groundwork, however, for this type of technology began over a century ago. Deep fakes are related to photo manipulation, which was developed as early as the 19th century. The technology, which could also be applied to motion pictures, improved throughout the 20th century.
The technology needed for modern-day deep fakes also involves a combination of software algorithms and AI. The specific technology needed for deep fakes was developed by a variety of researchers in the early 1990s. These developments occurred primarily in academic institutions. Unlike many other technological developments, there isn’t a single person or date on which it was developed.
There were, however, a few specific milestones in the creation of this technology. One such milestone was the Video Rewrite Program in 1997. This included software that was able to modify a video by placing an audio overlay on top of an existing video. This was the first type of software that could basically make a person say anything those creating the video desired. Since machine learning was the basis for how the software worked, this could be considered the time when deep fakes first started.
What Are the Application of Deep Fakes?
Deep Fakes are used in a variety of ways. Some of the positive ways are the following:
- Arts: Comedy and parody can be enhanced through the use of deep fake technology. Deep learning imagery could also be used in the gaming industry.
- Education: Deep fakes can help educators provide engaging lessons for students. A specific example is bringing historical figures to life in the classroom.
- Human Rights Activism: A deep fake might be used in order to disguise the voice and face of an individual who was working against a repressive regime. Journalists and activities can use this technology to mask their identity while reporting atrocities.
- Criminal Investigation: Artificial intelligence can now be used to reconstruct a crime scene. Using cell phone videos investigators can create a virtual crime scene.
Some of the negatives include the following:
- False Information: False information is sometimes used to sow discord in communities and throughout the country. This is done by creating deep fakes that show politicians and other prominent individuals saying and doing things that they never did.
- False Videos: The most common reason to make false videos is to put people’s faces and heads in pornographic movies. These attacks are often for revenge or extortion. Creating fake pornographic videos is a crime that has risen dramatically in recent years.
- Corporate Fraud: Criminals can create a deep fake video or audio of the company CEO and convince employees to transfer money to different bank accounts. This type of crime can cause major security threats to nearly every type of organization or company.
- Identity Theft: Using deep fakes of other individuals can be a form of identity theft if the other individual has not given their consent.
Examples of Deep Fakes In the Real World
The following are a few specific examples of deep fakes in recent years.
- Tom Cruise seemed to appear in a Tik-Tok video in early 2021. In the video Cruise was showing off coin tricks. This was a deep fake put together by Chris Umé and actor Miles Fisher.
- A young Paul McCartney was seen in a video that featured the song Find My Way. At the end of the video, the actor pulls off a “mask” to reveal he’s not really McCartney.
- Obama seemed to be making a public service announcement in a deep fake of the former president. Comedian Jordan Peele, using a variety of technology including fake apps, had his mouth pasted over Obama’s.
- Donald Trump appeared in a Breaking Bad segment as attorney Saul Goodman.
The risks from deep fakes can be reduced if people are aware of how prevalent they are and the ways they can sometimes be spotted. There are a few tips to follow that can help people more easily spot deep fakes.
In 2018, researchers found that faces in deep fakes don’t blink normally. At that point, the algorithms had not learned about blinking. This initially seemed like the best way to always tell a deep fake from the real thing. Not long after the research had been published, however, the deep fakes started showing faces that were blinking normally.
Spotting a deep fake is difficult because as soon as an area is revealed that helps people to notice a deep fake, the area is then fixed. There are still some ways to spot a deep fake, especially one that is of poor quality. Skin tone is sometimes patchy and hair doesn’t always look real in a fake video. Teeth and even jewelry might also look bad in a fake.